Density-Tempered Marginalized Sequential Monte Carlo Samplers

Abstract:     
In this paper we propose a new class of samplers for full Bayesian inference in general state-space models, called marginalized sequential monte carlo samplers. Here the dynamic states are approximately marginalized out using particle filters. Then, to sample from the fixed parameters, we set up a density-tempered bridge between the prior and the posterior and operate a sequential monte carlo sampler over this sequence. The approach delivers exact draws from the joint posterior of the fixed parameters and the hidden states for any given number of state particles and is easily parallelized. Given that it incorporates the sample information in a smooth fashion, it shows good performance in the presence of outliers. Further we also show how to use density tempering to automatically adjust the number of state particles. We illustrate the use of the method on a linear gaussian state space model, on a GARCH-type model of stock prices with microstructure noise and on a structural credit risk model with stochastic asset volatility.

Number of Pages in PDF File: 45

Keywords: Particle Filter, MCMC, Sequential Monte Carlo Samplers, Bayesian Methods

JEL Classification: C11

working papers series